Towards advanced mineral identification for future space mining applications employing LIBS and machine learning

Abstract

The growing interest in sustainable space exploration has brought in situ resource utilization (ISRU) to the forefront of planetary science. This study presents an integrated approach to autonomous mineral identification for space mining by combining Laser-Induced Breakdown Spectroscopy (LIBS) with supervised machine learning (ML). A dataset of over 400 high-resolution LIBS spectra representing 25 mineral classes was collected under simulated low-pressure conditions to replicate extraterrestrial environments. The raw spectra were preprocessed using wavelet-based denoising to reduce random noise, baseline correction to remove the background continuum, and spectral normalization to account for intensity variations. To simplify the data and enhance classification performance, three feature selection methods were applied: Principal Component Analysis (PCA), which identifies directions of maximum variance to reduce data dimensionality; variance thresholding, which removes spectral features with negligible variability across samples; and random forest-based feature selection (RF-FS), which ranks wavelengths by their importance for classification. Several classification algorithms were evaluated, with test accuracies reaching up to 89.3%. The best results were achieved using random forest and logistic regression models trained on features selected by RF-FS, showing strong generalization to previously unseen samples. This work demonstrates the potential of LIBS-ML integration for fast, robust, and accurate mineral classification, including reliable identification of dominant phases in mineral mixtures in planetary environments. The approach also provides interpretability and classifier confidence estimation, supporting adaptive autonomous mineral identification for future robotic exploration missions.